Purchase this article with an account.
A.O. Edwards, J. Yu, A. Khotanzad; Retinal image registration using thin–plate spline and vessel matching for multimodal studies on early detection and longitudinal monitoring of AMD . Invest. Ophthalmol. Vis. Sci. 2004;45(13):2988.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: Early detection and quantitative monitoring of early features of AMD such as drusen will require precise registration of retinal images obtained by different modalities. The gold–standard for studying drusen is color fundus photography that poorly penetrates the retinal pigment epithelium (RPE). IR achieves excellent penetration of the RPE leading to greater resolution of subretinal structures including drusen and detection at a smaller, earlier stage. However, the relationship between drusen features on IR and those seen on color photography is unknown. To study this relationship, exact matches of a location on one of the images should be found on the other image. However, the two images are taken by different devices and have translation, rotational and scale difference as well as spatial warping. Methods: We propose a two–stage registration method with (i) landmark–based thin–plate spline (TPS) interpolation to correct global spatial difference and (ii) vessel matching approach to improve local registration. To register images I(x,y) and J(u,v), we must find a mapping M such that for (u,v)=M(x,y), (x,y) and (u,v) are the same anatomic location. If known matches referred to as control points or landmarks (( xi,yi) and (ui,vi) ) are available, M can be decided through the TPS approach. Results: The TPS transformation works well for correcting global warping. However, manual selection of the landmarks introduces other errors and we use vessel matching to improve the registration. A modified Hausdorff distance proposed by Huttenlocher is used to quantify the accuracy of registration. For point set A and B, define Hε(A,B) = min(hε(A,B),hε(B,A)) and hε(A,B) = (#(ABε))/(#(A), where Bε is B dilated by ε, #(A) is the number of points in A, and AB is the intersection of A and B. For (xi,yi), the manually selected corresponding point is (ui,vi) whereas the perfect match is (Ui,Vi). A window W centered at (ui,vi) is created in image SI, which is the TPS corrected image of image I. The vessel–like pixels are then identified in this window and coded as binary (vessel, non–vessel) pixels. Then, a search in a local area around this point in image J is performed to find the best match to vessel pixels in W. The match is decided using the modified Hausdorff distance operating on vessel and non–vessel pixels. The pair consisting of (xi,yi) and the refined (Ui,Vi) is used to recalculate the TPS interpolation. Conclusions: The proposed registration scheme has been tested on color and IR image pairs with success.
This PDF is available to Subscribers Only